| Message ID | mpt8soalg7n.fsf@arm.com |
|---|---|
| State | New |
| Headers | show |
| Series | Reject versioning for alignment with different masks (PR 92526) | expand |
On Wed, Nov 20, 2019 at 4:28 PM Richard Sandiford <richard.sandiford@arm.com> wrote: > > Allowing mixed vector sizes broke the assumption in the following assert, > since it's now possible for different accesses to require different > levels of alignment: > > /* FORNOW: use the same mask to test all potentially unaligned > references in the loop. The vectorizer currently supports > a single vector size, see the reference to > GET_MODE_NUNITS (TYPE_MODE (vectype)) where the > vectorization factor is computed. */ > gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo) > || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask); > > I guess we could try to over-align smaller accesses so that all > of them are consistent, or try to support multiple alignment masks, > but for now the easiest fix seems to be to turn the assert into a > bail-out check. > > Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install? OK. > Richard > > > 2019-11-20 Richard Sandiford <richard.sandiford@arm.com> > > gcc/ > PR tree-optimization/92526 > * tree-vect-data-refs.c (vect_enhance_data_refs_alignment): Reject > versioning for alignment if the accesses do not have a consistent > mask, rather than asserting that the masks are consistent. > > gcc/testsuite/ > PR tree-optimization/92526 > * gcc.target/aarch64/pr92526.c: New test. > > Index: gcc/tree-vect-data-refs.c > =================================================================== > --- gcc/tree-vect-data-refs.c 2019-11-16 11:40:19.105159717 +0000 > +++ gcc/tree-vect-data-refs.c 2019-11-20 15:27:49.385346722 +0000 > @@ -2266,13 +2266,15 @@ vect_enhance_data_refs_alignment (loop_v > mask must be 15 = 0xf. */ > mask = size - 1; > > - /* FORNOW: use the same mask to test all potentially unaligned > - references in the loop. The vectorizer currently supports > - a single vector size, see the reference to > - GET_MODE_NUNITS (TYPE_MODE (vectype)) where the > - vectorization factor is computed. */ > - gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo) > - || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask); > + /* FORNOW: use the same mask to test all potentially unaligned > + references in the loop. */ > + if (LOOP_VINFO_PTR_MASK (loop_vinfo) > + && LOOP_VINFO_PTR_MASK (loop_vinfo) != mask) > + { > + do_versioning = false; > + break; > + } > + > LOOP_VINFO_PTR_MASK (loop_vinfo) = mask; > LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).safe_push (stmt_info); > } > Index: gcc/testsuite/gcc.target/aarch64/pr92526.c > =================================================================== > --- /dev/null 2019-09-17 11:41:18.176664108 +0100 > +++ gcc/testsuite/gcc.target/aarch64/pr92526.c 2019-11-20 15:27:49.385346722 +0000 > @@ -0,0 +1,9 @@ > +/* { dg-options "-O3 -mstrict-align" } */ > + > +void > +f (unsigned int *restrict x, unsigned int *restrict y, > + unsigned char *restrict z, unsigned int n) > +{ > + for (unsigned int i = 0; i < n % 4; ++i) > + x[i] = x[i] + y[i] + z[i]; > +}
Index: gcc/tree-vect-data-refs.c =================================================================== --- gcc/tree-vect-data-refs.c 2019-11-16 11:40:19.105159717 +0000 +++ gcc/tree-vect-data-refs.c 2019-11-20 15:27:49.385346722 +0000 @@ -2266,13 +2266,15 @@ vect_enhance_data_refs_alignment (loop_v mask must be 15 = 0xf. */ mask = size - 1; - /* FORNOW: use the same mask to test all potentially unaligned - references in the loop. The vectorizer currently supports - a single vector size, see the reference to - GET_MODE_NUNITS (TYPE_MODE (vectype)) where the - vectorization factor is computed. */ - gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo) - || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask); + /* FORNOW: use the same mask to test all potentially unaligned + references in the loop. */ + if (LOOP_VINFO_PTR_MASK (loop_vinfo) + && LOOP_VINFO_PTR_MASK (loop_vinfo) != mask) + { + do_versioning = false; + break; + } + LOOP_VINFO_PTR_MASK (loop_vinfo) = mask; LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo).safe_push (stmt_info); } Index: gcc/testsuite/gcc.target/aarch64/pr92526.c =================================================================== --- /dev/null 2019-09-17 11:41:18.176664108 +0100 +++ gcc/testsuite/gcc.target/aarch64/pr92526.c 2019-11-20 15:27:49.385346722 +0000 @@ -0,0 +1,9 @@ +/* { dg-options "-O3 -mstrict-align" } */ + +void +f (unsigned int *restrict x, unsigned int *restrict y, + unsigned char *restrict z, unsigned int n) +{ + for (unsigned int i = 0; i < n % 4; ++i) + x[i] = x[i] + y[i] + z[i]; +}
Allowing mixed vector sizes broke the assumption in the following assert, since it's now possible for different accesses to require different levels of alignment: /* FORNOW: use the same mask to test all potentially unaligned references in the loop. The vectorizer currently supports a single vector size, see the reference to GET_MODE_NUNITS (TYPE_MODE (vectype)) where the vectorization factor is computed. */ gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo) || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask); I guess we could try to over-align smaller accesses so that all of them are consistent, or try to support multiple alignment masks, but for now the easiest fix seems to be to turn the assert into a bail-out check. Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install? Richard 2019-11-20 Richard Sandiford <richard.sandiford@arm.com> gcc/ PR tree-optimization/92526 * tree-vect-data-refs.c (vect_enhance_data_refs_alignment): Reject versioning for alignment if the accesses do not have a consistent mask, rather than asserting that the masks are consistent. gcc/testsuite/ PR tree-optimization/92526 * gcc.target/aarch64/pr92526.c: New test.